Tomato Leaf Disease Detection
(2025)AI-Powered Crop Health Monitoring
Duration: Apr–May 2025
Synopsis
Group project: a custom YOLO11m detector, trained from scratch with no COCO pretraining, that finds and localises 7 tomato leaf disease classes. A Fourier-based loss keeps the fine lesion edges the network would otherwise blur away.
The Making Of
🎯 The Challenge
Detect and localise lesions across 7 disease classes from ~600 annotated images, without using an off-the-shelf pretrained detector, while staying light enough to run on simple hardware in the field.
💡 The Solution
Built a custom YOLO11m from scratch: two-conv C2f fusion blocks replacing three-conv blocks, SPPF for multi-scale context, and polarized self-attention before the detection head. Added a Fourier-domain loss that compares the high-frequency spectrum of each lesion with the backbone's features, penalising the network for losing edges.
🚀 The Impact
36% fewer parameters (28.2M vs ~44M) and ~27% faster inference than off-the-shelf YOLO11m with no accuracy loss. The Fourier loss lifted mAP@50 from 80.9% to 83.3% and F1 from 0.709 to 0.735, at zero inference cost.
Key Metrics
Tech Stack (Cast & Crew)
- YOLO-11MObject Detection
- PyTorchML Framework
- FFTFourier Loss
- PythonImplementation
Quick Info
- Year
- 2025
- Duration
- Apr–May 2025
- Complexity
- Senior
- Featured
- Yes